social recommendation: a review - ecology labfaculty.cs.tamu.edu/xiahu/papers/snam13.pdf · social...

Noname manuscript No.(will be inserted by the editor)

Social Recommendation: A Review

Jiliang Tang · Xia Hu · Huan Liu

Received: date / Accepted: date

Abstract Recommender systems play an important role in helping online usersfind relevant information by suggesting information of potential interest to them.Due to the potential value of social relations in recommender systems, social rec-ommendation has attracted increasing attention in recent years. In this paper, wepresent a review of existing recommender systems and discuss some research direc-tions. We begin by giving formal definitions of social recommendation and discussthe unique property of social recommendation and its implications compared withthose of traditional recommender systems. Then, we classify existing social recom-mender systems into memory-based social recommender systems and model-basedsocial recommender systems, according to the basic models adopted to build thesystems, and review representative systems for each category. We also presentsome key findings from both positive and negative experiences in building socialrecommender systems, and research directions to improve social recommendationcapabilities.

Keywords Social Recommendation · Social Recommender Systems · Recom-mender Systems · Social Network Analysis · Social Media

1 Introduction

With the development of the World Wide Web, information has increased at anunprecedented rate and the information overload problem has become increas-ingly severe for online users. For example, when we want to buy a computer and

Jiliang TangComputer Science, Arizona State UniversityE-mail: [email protected]

Xia HuComputer Science, Arizona State UniversityE-mail: [email protected]

Huan LiuComputer Science, Arizona State UniversityE-mail: [email protected]

2 Jiliang Tang et al.

search “computer” in Amazon, it returns 11,100,260 products1. Recommendersystems, which attempt to tackle the information overload problem by suggestinginformation that is of potential interest to online users, have become importantand popular [97,33,75,58,68]. For information customers, good recommendationsallow them to quickly find relevant information buried in a large amount of ir-relevant information. For information providers, recommender systems not onlyhelp determine which information to offer to individual consumers, but also im-prove consumer loyalty because consumers tend to return to the sites that bestserve their needs [99]. Such systems are widely implemented in various domainsincluding, product recommendation on Amazon2 and movie recommendation onNetflix3.

Recommender systems became an independent research area in the mid-1990s [3]and have attracted much attention from multiple disciplines, such as mathematics,physics, psychology, and computer science [26]. For example, the winners of theNetflix prize contest, one of the most famous competitions for recommendation,consist of psychologists, computer scientists, and physicists4. Many techniquesare used to build recommender systems, which can be generally classified intocontent-based methods, collaborative filtering (CF) based methods, and hybridmethods [3]. Content-based methods, rooted in information retrieval [7] and in-formation filtering research [9], recommend items similar to the ones the user pre-ferred in the past; CF-based methods predict user interests directly by uncoveringcomplex and unexpected patterns from a user’s past behaviors and recommenditems to a user from other users with similar interests and preferences in thepast [57,105]. Hybrid methods combine content-based and CF-based methods. Wewill give a brief review about techniques to build recommender systems in Section2.

The increasing popularity of social media greatly enriches people’s social activi-ties with their families, friends, and colleagues, which produces rich social relationssuch as friendships in Facebook5, following relations in Twitter6 and trust rela-tions in Epinions7. Online social relations provide a different way for individualsto communicate digitally and allow online users to share ideas and opinions withtheir connected users. A user’s preference is similar to or influenced by their so-cially connected friends and the rationale behind the assumption can be explainedby social correlation theories such as homophily [77] and social influence [71].Homophily indicates that users with similar preferences are more likely to be con-nected, and social influence reveals that users who are connected are more likely tohave similar preferences. Analogous to the fact that users in the physical world arelikely to seek suggestions from their friends before making a purchase decision andusers’ friends consistently provide good recommendations [103], social relationscan be potentially exploited to improve the performance of online recommendersystems [33,67,49,68].

1 http://www.amazon.com/s/ref=nb sb noss 1?url=search-alias%3Daps&field-keywords=computer

2 http://www.amazon.com/3 http://www.netflix.com4 http://en.wikipedia.org/wiki/Netflix Prize5 https://www.facebook.com/6 https://www.twitter.com/7 http://www.epinions.com

Social Recommendation: A Review 3

In this article, we present a review of existing social recommender systemsand discuss some research directions that can improve the capabilities of socialrecommendation. The motivation of this article is multi-fold:

– Social recommendation has been developed for several years and still there is noagreement in literature on the definition of social recommendation. These clar-ifications are necessary for people to gain an awareness of social recommendersystems. For example, Kaifu Lee, CEO at Innovation Works, recommended aparticular brand of frozen dessert maker though Chinese tweet Sina Weibo plat-form and a few hours later, Taobao (China’s eBay) had hundreds of sellers ofthe item and thousands of the machines were sold within a day. Subsequently,he wrote an article in LinkedIn entitled “The Power of Social Recommenda-tion”, but there are many doubts about whether this kind of recommendationcan be called social recommendation8.

– Social relations provide an independent source for recommendation; variousapproaches are proposed to build social recommender systems such as trustensemble [70], trust propagation [49], and social regularization [68]. A clas-sification of existing social recommender systems can help users gain a quickunderstanding of many social recommender systems. A review of existing socialrecommender systems will let users familiarize themselves with the state-of-the-art social recommender systems.

– On the one hand, there is recent work reporting significant recommendationperformance improvement for social recommender systems [33,67,49,68,52].On the other hand, there are also unsuccessful attempts at applying socialrecommendation [15,30,92,48]. For example, IBM’s Black Friday report saysTwitter delivered 0 percent of referral traffic and Facebook sent just 0.68 per-cent [48]. It is important to discuss key findings from both positive and negativeexperiences to seek a deeper understanding and further development of socialrecommender systems.

– Social recommendation is still in the early stages of development, and thereare many challenging issues needing further investigation. It is necessary todiscuss some research directions that can improve social recommendation ca-pabilities and make social recommendation applicable to an even broader rangeof applications.

Our major contributions are summarized below:

– We give narrow and broad definitions of social recommendation to cover mostexisting definitions of social recommendation in the literature and discuss theunique property of social recommender systems and its implications comparedwith traditional recommender systems;

– We present a classification of social recommender systems according to thebasic models to build social recommender systems, and review representativesystems for each category in detail;

– We summarize key findings from some positive and negative experiences inapplying social recommender systems;

– We discuss some research directions to improve recommendation capabilitiesincluding exploiting the heterogeneity of social networks and weak dependence

8 http://www.linkedin.com/today/post/article/20121203134252-416648-the-power-of-social-recommendation


connections, segmenting users and items, considering the temporal informationof rating and social information, understanding the role of negative relations,and integrating cross-media social networks.

The rest of this article is organized as follows. We review existing recommendersystems including content-based methods, collaborative filtering based methodsand hybrid methods in Section 2. In Section 3, we formally define social recom-mendation, discuss the unique property of social recommendation, classify andreview existing social recommender systems, and discuss some key findings fromboth positive and negative experiences in applying social recommender systems.In Section 4, we discuss possible research directions for improvement of socialrecommendation. The article is summarized in Section 5.

2 Traditional Recommender Systems

In a typical recommender system, there is a set of users and a set of items. Letu = u1, u2, . . . , un and v = v1, v2, . . . , vm be the sets of users and itemsrespectively, where n is the number of users and m is the number of items. A userui rates a subset of items with some scores. We use R ∈ R

n×m to denote therating matrix where Rij is the rating score if ui gives a rating to vj , otherwise weemploy the symbol “?” to denote the unknown rating. Usually the rating matrix isvery sparse, suggesting that there are lots of unknown ratings in R. For example,the density of the rating matrix in commercial recommender systems is often lessthan 1% [97]. If item vj has attributes, we use xj ∈ R

ℓ to represent vj where ℓ

is the number of attributes. The task of recommender systems is to predict therating for user ui on a non-rated item vj or to recommend some items for givenusers, i.e., to predict missing values in R based on known ratings.

Since recommender systems first became an independent research area in themid-1990s, there have been many recommender systems proposed, which can begenerally grouped into content based, collaborative filtering (CF) based and hybridrecommender systems [95,51]. In the following subsection, we will briefly revieweach category with its representative systems.

2.1 Content Based Recommender Systems

Content based recommender systems have their roots in information retrieval [7]and information filtering research [9]. They recommend items similar to the onesthat the user has preferred in the past. Most existing content based recommendersystems focus on recommending items with textual information such as news,books and documents. The content in these systems is usually described withkeywords [8,90] and the informativeness of a keyword to a document is oftenmeasured by TFIDF weight [121]. TF weight of a keyword to a document denotesthe frequency of the keyword in the document, while IDF weight of a keyword isdefined as the inverse document frequency of the keyword.

Let xjk be the TFIDF weight of the k-th keyword in vj and the content ofvj can be represented as xj = (xj1, xj2, . . . , xjℓ). With these representations,content based recommender systems recommend items to a user which are similarto items the user liked in the past [90]. In particular, various candidate items are


compared with items previously rated by the user. A similarity measure such asa cosine similarity measure is adopted to score these candidate items. Besides thetraditional heuristics based information retrieval methods, there are also contentbased recommender systems that use other techniques such as various classificationand clustering algorithms [90,82].

As noticed in [8,3], content based recommender systems have several limita-tions: (1) limited content analysis - these systems are difficult to apply to domainswhich have an inherent problem with automatic feature extraction such as multi-media data; (2) over-specialization - items recommended to a user are limited tothose similar to items the user already rated; (3) new user problem - to let contentbased recommender systems understand a user’s preference, the user has to ratea sufficient number of items, hence, content based recommender systems fail torecommend items for users with few or no ratings.

2.2 Collaborative Filtering Based Recommender Systems

Collaborative filtering is one of the most popular techniques to build recommendersystems [97,57,105]. It can predict user interests directly by uncovering complexand unexpected patterns from a user’s past behaviors such as product ratings with-out any domain knowledge [57,105]. The underlying assumption of collaborativefiltering based recommender systems is that if users have agreed with each otherin the past, they are more likely to agree with each other in the future than toagree with randomly chosen users. Existing collaborative filtering methods can becategorized into memory-based methods and model-based methods [36,10,105].

2.2.1 Memory based Collaborative Filtering

Memory based methods use either the whole user-item matrix or a sample togenerate a prediction [105], which can be further divided into user-oriented meth-ods [44,10] and item-oriented methods [97,53]. User-oriented methods predict anunknown rating from a user on an item as the weighted average of all the ratingsfrom her similar users on the item, while item-oriented methods predict the ratingfrom a user on an item based on the average ratings of similar items by the sameuser. The key problems a memory-based CF method has to solve are comput-ing similarity and aggregating ratings. User-oriented methods and item-orientedmethods can leverage similar techniques to address these two problems; thus, weuse user-oriented methods as examples to illustrate representative methods forcomputing similarity and aggregating ratings.

Computing similarity for user-oriented methods : Computing user-user sim-ilarity is a critical step for user-oriented methods. There are many techniquesproposed to tackle this problem such as Pearson Correlation Coefficient [94], Co-sine similarity [17], and probability-based similarity [53,22], among which PearsonCorrelation Coefficient and Cosine similarity are the most widely used ones.

– Pearson Correlation Coefficient : Each user is presented as a vector of ratings.For example, the i-th user will be denoted as Ri. Pearson Correlation Co-efficient measures the extent to which two variables linearly relate with each


other [94]. Pearson Correlation Coefficient between ui and uj can be calculatedas

Sij =

∑

k∈I(Rik − Ri) · (Rjk − Rj)√

∑

k∈I(Rik − Ri)2√

∑

k∈I(Rjk − Rj)2, (1)

where I denotes the set of items rated by both ui and uj and S ∈ Rn×n

represents the user-user similarity matrix. Ri denotes the average rate of ui.– Cosine similarity : Cosine similarity computes the cosine of the angle formed

by the rating vectors [17]. For example, the cosine similarity between ui anduj can be calculated as,

Sij =

∑

k∈I Rik ·Rjk√

∑

k∈I R2ik

√

∑

k R2jk

, (2)

Aggregating ratings for user-oriented methods : After obtaining a user-user sim-ilarity matrix, user-oriented methods will predict a missing rating for a given userby aggregating the ratings of users similar to her. Various aggregating strategieshave been proposed [94,97,53] and the most widely used strategy is weighted av-erage rating as

Rij = Ri +

∑

uk∈NiSik(Rkj − Rk)

∑

uk∈NiSik

, (3)

where Ni is the set of users who have rated the j-th item vj .

2.2.2 Model based Collaborative Filtering

Model-based methods assume a model to generate the ratings and apply data min-ing and machine learning techniques to find patterns from training data [124,105],which can be used to make predictions for unknown ratings. Compared to memorybased CF, Model-based CF has a more holistic goal to uncover latent factors thatexplain observed ratings [124]. Well-known model-based methods include Bayesianbelief net CF models [10,81], clustering CF models [115,11], random walk basedmethods [124,47] and factorization based CF models [37,45,89,57,96]. Factoriza-tion based CF methods [57,96] are very competitive if not the best and are widelyadopted to build recommender systems [12].

Factorization based CF models assume that a few latent patterns influenceuser rating behaviors and perform a low-rank matrix factorization on the user-item rating matrix. Let ui ∈ R

K and vj ∈ RK be the user preference vector

for ui and item characteristic vector for vj respectively, where K is the number oflatent factors. Factorization based collaborative filtering models solve the followingproblem

minU,V

n∑

i=1

m∑

j=1

Wij(Rij −UiV⊤j )2 + α(‖U‖2F + ‖V‖2F ), (4)

where U = [U⊤1 ,U⊤

2 , . . . ,U⊤n ]⊤ ∈ R

n×k andV = [V⊤

1 ,V⊤2 , . . . ,V⊤

m]⊤ ∈ Rm×K . K is the number of latent factors (patterns),


which is usually determined via cross-validation. The term α(‖U‖2F + ‖V‖2F ) isintroduced to avoid over-fitting, controlled by the parameter α. W ∈ R

n×m isa weight matrix where Wij is the weight for the rating for ui to vj . A commonway to set W is Wij = 1 if Rij 6= 0. The weight matrix W can also be usedto handle the implicit feedback and encode side information such as user clickbehaviors [28], similarity between users and items [88,62], quality of reviews [93]and user reputation [112].

Collaborative filtering based recommender systems can overcome some of theshortcomings of content based recommender systems. For example, CF based sys-tems use rating information; hence, they are domain-independent and can recom-mend any items. However, CF based systems have their own limitations such asthe cold-start problem (new items or new users) and the data sparsity problem [3,105].

2.3 Hybrid Recommender Systems

To avoid certain limitations of content and collaborative filtering systems, hybridapproaches combine content and CF based methods. Various strategies are pro-posed to combine content and CF based methods, which can be roughly classifiedinto three categories [3]. We will briefly review each category with representativesystems.

Combining different recommenders: under this strategy, content and CF basedmethods are implemented separately and then their predictions are combined toobtain the final recommendation. Various ways are proposed, such as a votingscheme [91] and a linear combination of ratings [18], to combine predictions fromcontent and CF based methods.

Adding content based characteristics to CF models: systems with this strategyuse content based profiles and uncommonly rated items to calculate user-usersimilarities. These systems can overcome some sparsity-related problems of CFmethods and recommend items directly when item scores highly against the user’sprofiles [91,38].

Adding CF based characteristics to content based models: the most popularapproach under this strategy is to use a dimensionality reduction technique on thecontent profile matrix. For example, in [104] latent semantic indexing is used tocreate a collaborative view of a set of user profiles, which improves recommendationperformance compared to pure content based approaches.

3 Social Recommendation

One of the earliest social recommender systems appeared in 1997 [54]. Myriads ofsocial media services such as Facebook and Twitter have emerged in recent years toallow people to easily communicate and express themselves conveniently. The per-vasive use of social media generates social information at an unprecedented rate.For example, Facebook, the largest social networking site produces 35,000,000,000online friendships [41] and the most popular user on Twitter, the largest microblog-ging site, has 37, 974, 138 followers9. The rapid development of social media has

9 http://twitaholic.com/


greatly accelerated the development of social recommender systems [55,41]. Inthis section, we will first give the definitions of social recommendation, discussthe opportunities for social recommendation systems compared to traditional rec-ommender systems, classify and review existing social recommender systems, andfinally summarize some key findings from positive and negative experiences inapplying social recommender systems.

3.1 Definitions of Social Recommendation

Social recommendation has been studied since 1997 [54] and has attracted increas-ing attention with the growing popularity of social media [55,41], however, it hasno commonly accepted definition. In this subsection, we give narrow and broaddefinitions of social recommendation based on existing definitions in the literature.

A narrow definition of social recommendation is any recommendation withonline social relations as an additional input, i.e., augmenting an existing rec-ommendation engine with additional social signals. Social relations can be trustrelations, friendships, memberships or following relations. A tutorial with the ti-tle “Introduction to Social Recommendation” at the 19th international world wideweb conference (WWW2010) followed this narrow definition [55]. In this definition,social recommender systems assume that users are correlated when they establishsocial relations [72,75,67]. For example, users’ preferences are likely to be similarto or influenced by their connected friends. Under this assumption, social recom-mendation leverages user correlations implied by social relations to improve theperformance of recommendation. Representative systems under this definition in-clude TidalTrust [34], MoleTrust [75,118], SoRec [70], SocialMF [49], SoReg [68]and LOCALBAL [112].

Another tutorial titled “Social Recommender Systems” at the 20th interna-tional world wide web conference (WWW2011) adopted a broad definition of so-cial recommendation where social recommender systems are defined as any rec-ommender systems that target social media domains [41]. The definition coversrecommender systems recommending any objects in social media domains such asitems (the focus of recommender systems under the narrow definition), tags [102],people [13,5], and communities [14]. The sources they use are not limited to onlinesocial relations but include all kinds of available social media data such as socialtagging [69], user interactions [52] and user click behaviors [78].

With narrow and broad definitions, let us go back to examine the example ofKaifu Lee’s recommendation. Since the recommendation happens in social mediadomains, Kaifu Lee’s recommendation is a sort of social recommendation under thebroad definition. However, the sources mentioned in this recommendation are notlimited to social networks, and this recommendation does not satisfy the narrowdefinition.

In this article, we focus on social recommender systems under the narrow def-inition. The reason is two fold. First, similar techniques can be applied to imple-menting social recommender systems under both definitions. Second, the narrowdefinition is straightforward and helps readers gain a deeper understanding ofstate-of-the-art social recommender systems.


3.2 A Unique Feature of Social Recommendation and Its Implications

The increasing popularity of social media allows online users to participate in on-line activities which produce rich social relations. In social recommender systems,in addition to the rating information R, users can connect to each other. LetT ∈ R

n×n denote user-user social relations where Tij = 1 if uj connects to ui

and zero otherwise. The availability of social relations T provides an independentsource for recommendation and this unique property of social recommendationbrings about new opportunities.

First, traditional recommender systems assume that users are independent andidentically distributed (known as the i.i.d. assumption). However, online users areinherently connected via various types of relations such as friendships and trustrelations. Figure 1 shows a simple example of connected users in social media,and Figures 1(b) and 1(c) demonstrate its two data representations for traditionalrecommender systems and social recommender systems, respectively. In additionto the rating matrix in traditional recommender systems, users in social recom-mender systems are connected, providing social information. Since they are con-nected, users are correlated rather than independent and identically distributed.For example, [120] finds that users with following relations are more likely to sharesimilar interests in topics than two randomly chosen users, and [109] shows thatusers with trust relations are more likely to have similar preferences in item ratings.This phenomenon is observed in most online social networks and can be explainedby social correlation theories such as homophily [77] and social influence [71]. Notethat one user is more likely to have similar interests with her connected users thanthose randomly chosen users, which does not indicate that her connected usersare equivalent to her most similar users w.r.t. rating information. Assume that ui

connects to d users. We use Ci and Si with the size d to denote sets of the con-nected users and the top-d similar users of ui, respectively. In [108], the authorsshow that the overlap between Ci and Si is less than 10%, which is consistent withthe observation in [19]. If we change our perspective and consider connections assimilarity measurements, social information provides similarity evidence and Ci isthe list of the similar users w.r.t. social information [43]. Then the above observa-tion can be explained by the key finding in [42] - lists of similar users obtained bydifferent sources are found to be highly different from each other. It also suggeststhat an aggregation of rating and social information may be valuable. In additionto similarity evidence, social information also provides another unique evidence,i.e., familiarity evidence, which is very importance in recommendation since in thephysical world, we usually ask suggestions from our friends who are familiar withour tastes [43,42]. Both similarity and familiarity evidence from social informa-tion indicate that social networks contain complementary information to ratinginformation and provide an independent source of information about online users.Exploiting social relations can potentially improve recommendation performance.

Second, besides recommender systems, social recommendation involves the in-dependent research field of social network analysis [119,100,101,20]. Social networkanalysis (SNA) is the methodical analysis of social networks and has emerged as akey technique in modern sociology. It has gained a significant following in variousdisciplines such as communication studies, economics, geography, and computerscience; consumer tools for it are now commonly available [101]. Social recom-mender systems can take advantage of research results from social network anal-


!

"

#

$

$

$"

$#

$!

5

2 4

4

5

1

4

4

1

3

(a) Connected Users

5 ? 2 ? ?

4 4 5 ? ?

? ? 4 4 1

? ? ? ? 3

? ? 1 ? ?

!

"

#

$ $ $! $" $#

(b) Traditional Recommender Systems

0 1 0 0 1

0 0 1 1 0

0 0 0 0 0

1 0 1 0 0

0 1 1 1 0

!

"

#

! " #

5 ? 2 ? ?

4 4 5 ? ?

? ? 4 4 1

? ? ? ? 3

? ? 1 ? ?

!

"

#

$ $ $! $" $#

(c) Social Recommender Systems

Fig. 1 A Simple Example of Connected Users


ysis such as social correlation theories [77,71], status analysis [87,56], communitydetection [60,114], online trust [72,110] and heterogeneous networks [106] for rec-ommendation. Research achievements from social network analysis pave ways toexploit social relations and can be applied to building social recommender systems.

3.3 Existing Social Recommender Systems

As mentioned earlier, collaborative filtering is widely adopted to build recom-mender systems, and most existing social recommender systems are based on CFtechniques. Therefore, our discussion in this article is focused on CF based socialrecommender systems. Social recommendation has two inputs, i.e., rating infor-mation and social information, as shown in Figure 1(c). Most existing social rec-ommender systems choose CF models as their basic models to build systems andpropose approaches to capture social information based on results from social net-work analysis. Therefore, a general CF based social recommendation frameworkcontains two parts: (1) a basic CF model and (2) a social information model, whichcan be formally stated as

a social recommendation CF model =

a basic CF model + a social information model (5)

The basic CF model in a social recommendation CF model creates a wayto classify social recommender systems. Following the classification of CF basedrecommender systems, we classify social recommender systems into two majorcategories according to their basic CF models: memory based social recommendersystems and model based social recommender systems.

3.3.1 Memory based Social Recommender Systems

Memory based social recommender systems use memory based CF models, espe-cially user-oriented methods as their basic models. A missing rating for a givenuser is aggregated from the ratings of users correlated to her, denoted as N+. Fora given user, traditional user-oriented methods use similar users N , while memorybased social recommender systems use correlated users N+ obtained from bothrating information and social information. Social recommender systems in this cat-egory usually follow two steps. First, they obtain the correlated users N+(i) for agiven user ui, and second is the classical last step of memory based CF methods- aggregating ratings from the correlated users obtained by the first step for miss-ing ratings. Different social recommender systems in this category use differentapproaches to obtain correlated users N+ in the first step; next, we will presentdetails about some representative approaches.

Social based Weight Mean [117,118]: For a given ui, this strategy simplyconsiders a ui’s directly connected users F(i) as the set of her correlated usersN+(i),

N+(i) = uj |T(i, j) = 1 (6)

TidalTrust [34]: Users in this system are connected via trust relations. Theauthors design a metric TidalTrust to estimate trust values among users based


on the following two observations: (1) shorter propagation paths produce moreaccurate trust estimates; (2) paths with higher trust values create better results.To estimate trust values among users, TidalTrust performs the following steps,

– searching a shortest path from the source user to raters and setting the shortestpath length as the path depth of the algorithm;

– computing trust value from the source user to a rater at the given depth. Fora pair of users ui and uk who are not directly connected, a trust value isaggregated from the trust value from ui’s direct neighbors to uj , weighted bythe direct trust values from ui to her direct neighbors as

Sij =

∑

uk ∈ F(i)SikSkj∑

uk ∈ F(i)Sik

(7)

– after trust values S are calculated, N+(i) is defined as the set of users whosetrust values with ui exceeds a given threshold τ ,

N+(i) = uj |Sij ≥ τ (8)

MoleTrust [73–75]: A new trust metric, MoleTrust, is proposed and con-sists of two major steps. First, cycles in trust networks are removed. To obtaintrust values, a large number of trust propagations have to be executed. Therefore,removing trust cycles beforehand from trust networks can significantly speed upthe proposed algorithm because every user only needs to be visited once to infertrust values. With this operation, the original trust network is transformed into adirected acyclic graph. Second, trust values are calculated based on the obtaineddirected acyclic graph by performing a simple graph random walk: first the trust ofthe users at 1-hop away is computed, then the trust of the users at 2-hop away, etc.After trust values are computed, MoleTrust defines users within maximum-depthand have rated the target item as correlated users N+, where maximum-depth isa predefined parameter.

Although MoleTrust has similar operations to TidalTrust, they are very differ-ent in two aspects. First, the definitions of correlated users in these two systemsare different. TidalTrust adds a user uk to N+(i) only if she is on a shortest pathfrom ui with respect to a target item, while MoleTrust considers all users who haverated the target item and that can be reached through a direct or propagated trustrelation within maximum-depth as N+(i). Second, MoleTrust needs a predefinedtrust threshold to determine users who will be considered in the rating aggregat-ing process, while TidalTrust determines the trust threshold automatically, whichdenotes the path strength (the minimum trust rating on a path).

TrustWalker [49]: The intuition of this system is from two key observations.First, a user’s social network has little overlap with users similar to her [19],suggesting that social information provides an independent source of information.Second, ratings from strongly trusted friends on similar items are more reliablethan ratings from weakly trusted neighbors on the same target item. The firstobservation indicates the importance of trust-based approaches while the secondobservation suggests the capability of item-oriented approaches. To take advantageof both approaches, TrustWalker proposes a random walk model to combine trustbased and user oriented approaches into a coherent framework. It queries a user’sdirect and indirect friends’ ratings for the target item as well as similar items byperforming random walk in online social networks. For example, to obtain a rating


for ui to vj . Suppose that we are at a certain node uk. Then TrustWalker worksas follows at each step of a random work: if uk rated vj , then it stops the randomwalk and returns Rkj as the result of random walk; otherwise, it has two choices- (1) it also stops random walk and randomly selects one of the items vk similarto vj rated by ui and returns Rik, or (2) it continues random walk and walks toauthor user uk in ui’s trust networks.

TrustWalker employs the Pearson Correlation Coefficient of ratings expressedfor items to calculate item-item similarity. Since values of Pearson CorrelationCoefficient are in the range of [−1, 1], only items with positive correlation withthe target item are considered. The similarity between vi and vj is then computedas,

sim(i, j) =1

1 + e−Nij

2

× PCC(i, j) (9)

where Nij is the number of users who rated both vi and vj , and PCC(i, j) isPearson Correlation Coefficient of vi and vj .

3.3.2 Model based Social Recommender Systems

Model-based social recommender systems choose model-based CF methods as theirbasic models. Matrix factorization techniques are widely used in model basedCF methods. There are several nice properties of these matrix factorization tech-niques [25,79]: (1) many optimization methods such as gradient based methods canbe applied to find a well-worked optimal solution, scaled to thousands of users withmillions of trust relations; (2) matrix factorization has a nice probabilistic inter-pretation with Gaussian noise; (3) it is very flexible and allows us to include priorknowledge. Most existing social recommender systems in this category are basedon matrix factorization [67,65,126,50,116,68,6,123,107,110,111,86,46,109]. Thecommon rationale behind these methods is that users’ preferences are similar to orinfluenced by users whom they are socially connected to. However, the low cost ofsocial relation formation can lead to social relations with heterogeneous strengths(e.g., weak ties and strong ties mixed together) [122]. Since users with strong tiesare more likely to share similar tastes than those with weak ties, treating all socialrelations equally is likely to lead to degradation in recommendation performance.Therefore for each social relation, these methods associate a strength, which isusually calculated by rating similarity in existing social recommender systems.For example, when we choose cosine similarity, if ui and uk are connected , Sik iscalculated as the cosine similarity between the rating vectors of ui and uk, oth-erwise we set Sik to 0. Therefore, the strength between ui and uk Sik can beformally defined as

Sik =

∑jRij ·Rkj√∑

jR2

ij

√∑jR2

kj

when ui and uk are connected,

0 otherwise.(10)

Unlike traditional matrix factorization based recommender systems, social rec-ommender systems in this category can take advantage of social information, anda unified framework can be stated as,


minU,V,Ω

‖W ⊙ (R−U⊤V)‖2F + α Social(T,S, Ω)

+ λ(‖U‖2F + ‖V‖2F + ‖Ω‖2F ), (11)

where Social(T,S, Ω) is the term introduced to capture social information basedon social network analysis and Ω is the set of parameters learned from socialinformation. α is employed to control the contributions from social information. Wcontrols the weights of known ratings in the learning process. According to differentdefinitions of Social(T,S, Ω), we further divide social recommender systems inthis category into three groups: co-factorization methods, ensemble methods, andregularization methods. Next, we will review some representative systems in detailfor each group.

Co-factorization methods [67,109]: The underlying assumption of systemsin this group is that the i-th user ui should share the same user preference vector ui

in the rating space (rating information) and the social space (social information).Social recommender systems in this group perform a co-factorization in the user-item matrix and the user-user social relation matrix by sharing the same userpreference latent factor. SoRec [67] and LOCABAL [109] are two representativesystems in this group.

SoRec [67]: SoRec defines Social(T,S, Ω) as

minn∑

i=1

∑

uk∈Ni

(Sik − u⊤i zk)

2, (12)

where Ni is the set of users directly connected to ui and Z = z1, z2, . . . , zn ∈R

K×n is the factor-specific latent feature matrix. With this term, the user prefer-ence matrix U is learned from both rating information and social information bysolving the following optimization problem:

minU,V,Z

‖W ⊙ (R−U⊤V)‖2F + α

n∑

i=1

∑

uk∈Ni

(Sik − u⊤i zk)

2

+ λ(‖U‖2F + ‖V‖2F + ‖Z‖2F ) (13)

LOCABAL [109]: LOCABAL defines Social(T,S, Ω) as

min

n∑

i=1

∑

uk∈Ni

(Sik − u⊤i Huk)

2. (14)

LOCABAL is different from SoRec and it is based on social correlation theorieswhere the user preferences of two socially connected users are correlated via thecorrelation matrix H. In Eq. (14), for two socially connected users ui and uk, theirpreference vectors ui and uk are correlated through H, which is controlled by theirsocial strength Sik where H ∈ R

K×K is the matrix to capture the user preferencecorrelation. A large value of Sik, i.e., ui and uk with a strong connection, indicatesthat their preferences ui and uk should be tightly correlated via H, while a smallvalue of Sik indicates that ui and uk should be loosely correlated. Social(T,S, Ω)can be applied to both directed and undirected social networks via the correlation


matrix H. For a directed social network, the learned matrix H is asymmetric,while for an undirected social network, the learned matrix H is symmetric.

With the definition of Social(T,S, Ω), LOCABAL solves the following opti-mization problem,

minU,V,Z

‖W ⊙ (R−U⊤V)‖2F + α

n∑

i=1

∑

uk∈Ni

(Sik − u⊤i Huk)

2

+ λ(‖U‖2F + ‖V‖2F + ‖H‖2F ) (15)

Note that LOCABAL and SeRoc have the same form by further factorizing zkinto Huk. After obtaining factorized factors of the matrix S, we can use the factorsto reconstruct S. For example, for LOCABAL, after learning U and H, we canreconstruct S as S = U⊤HU. The reconstructed matrix S can be used to performsocial relation prediction [80,109]. Therefore, one advantage of approaches in thisgroup is that they jointly perform recommendation and social relation prediction.

Ensemble methods [65,110]: The basic idea of ensemble methods is thatusers and their social networks should have similar ratings on items, and a missingrating for a given user is predicted as a linear combination of ratings from the userand her social network. We give details about two representative systems below.

STE [65]: The rating from the i-th user ui to the j-th item vj will be estimatedby STE as

Rij = u⊤i vj + β

∑

uk∈Ni

Siku⊤k vj , (16)

where∑

uk∈NiSiku

⊤k vj is a weighted sum of the predicted ratings for vj from ui’s

social network, and β controls the influence from social information. It is easy toverify that Eq. (16) is equivalent to the following matrix form:

R = (I + βS)U⊤V. (17)

STE is to minimize the following term,

‖W ⊙ ((R−U⊤V)− βSU⊤V))‖2F= ‖W ⊙ (R−U⊤V)‖2F + αSocial(T,S, Ω) (18)

where Social(T,S, Ω) is defined as,

‖W ⊙ (R− βSU⊤V)‖2F− 2Tr(W ⊙ ((R−U⊤V)(βSU⊤V))) (19)

mTrust [110]: a variant of mTrust will predict the rating from ui to vj as,

Rij = u⊤i vj + β

∑

uk∈NiSikRkj

∑

uk∈NiSik

, (20)

where∑

uk∈NiSikRkj

∑uk∈Ni

Sikis a weighted mean of the ratings for vj from ui’s social

network. mTrust solves the following optimization problem,

minU,V,S

∑

i

∑

j

(Rij − u⊤i vj − β

∑

uk∈NiSikRkj

∑

uk∈NiSik

)2 (21)


Similar to STE, we can also find a standard form of Eq. (11) from Eq. (21) formTrust where Social(T,S, Ω) is defined as,

∑

i

∑

j

Wij((Rij − β

∑

uk∈NiSikRkj

∑

uk∈NiSik

)2

− 2(Rij − u⊤i vj)(β

∑

uk∈NiSikRkj

∑

uk∈NiSik

)) (22)

Note that there are two major differences between STE and mTrust. First,Sik denotes the influence from uk to ui for mTrust and will be learned from thedata automatically, while Sik in STE is the predefined similarity between ui anduk. Second, STE incorporates a weighted sum of the predicted ratings from socialnetworks, while mTrust incorporates a weighted mean of the existing ratings fromsocial networks.

Regularization methods [50,68]: Regularization methods focus on a user’spreference and assume that a user’s preference should be similar to that of hersocial network. For a given user ui, regularization methods force her preference ui

to be closer to that of users in ui’s social network Ni. SocialMF [50] and SocialRegularization [68] are two representative systems in this group.

SocialMF [50]: SocialMF forces the preference of a user to be closer to theaverage preference of the user’s social network and defines Social(T,S, Ω) as,

min

n∑

i=1

(ui −∑

uk∈Ni

Sikuk)2, (23)

where∑

uk∈NiSikuk is the weighted average preference of users in ui’s social

networkNi and SocialMF requires each row of S to be normalized to 1. The authorsdemonstrated that SocialMF addresses the transitivity of trust in trust networksbecause a user’s latent feature vector is dependent on the direct neighbors’ latentfeature vectors, which can propagate through the network and make a user’s latentfeature vector dependent on possibly all users in the network.

With the term to capture social information, SocialMF solves the followingoptimization problem,

minU,V

‖W ⊙ (R−U⊤V)‖2F + α

n∑

i=1

(ui −∑

uk∈Ni

Sikuk)2

+ λ(‖U‖2F + ‖V‖2F ) (24)

Social Regularization [68]: For a given user, users in her social network mayhave diverse tastes. With this intuition, social regularization proposes a pair-wiseregularization as,

min

n∑

i=1

∑

uk∈Ni

Sik(ui − uk)2, (25)

where the preference closeness of two connected users is controlled by their sim-ilarity based on their previous ratings. Similarity can be calculated by PearsonCorrelation Coefficient or Cosine similarity of commonly rated items by two con-nected users. A small value of Sik indicates that the distance between latent feature


vectors ui and uk should be larger, while a large value indicates that the distancebetween the latent feature vectors should be smaller. Eq. (25) can be rewritteninto its matrix form as,

1

2

n∑

i=1

∑

uk∈Ni

Sik(ui − uk)2

=n∑

i=1

n∑

k=1

Sik(ui − uk)2

=1

2

n∑

i=1

n∑

k=1

K∑

j=1

Sik

(

Uij −Ukj

)2

=n∑

i=1

n∑

k=1

K∑

j=1

SikU2ij

−n∑

i=1

n∑

k=1

K∑

j=1

SikUijUkj

= Tr(U⊤LU), (26)

where L = D − S is the Laplacian matrix and D is a diagonal matrix with thei-th diagonal element D(i, i) =

∑nj=1 S(j, i).

Recommender systems with social regularization solve the following problem,

minU,V

‖W ⊙ (R−U⊤V)‖2F + α

n∑

i=1

∑

uk∈Ni

Sik(ui − uk)2

+ λ(‖U‖2F + ‖V‖2F ) (27)

One advantage of approaches in this category is that they indirectly model thepropagation of tastes in social networks, which can be used to reduce cold-startusers and increase the coverage of items for recommendation.

3.4 Discussion

Social recommendation can potentially solve some challenging problems of tradi-tional recommender systems such as the data sparsity problem and the cold-startproblem, and has attracted broad attention from both academia and industry. Onthe one hand, social recommendation has been studied for many years in literature.Many successful systems have been proposed and recommendation performanceimprovement is reported. On the other hand, successful systems in industry arerare and there is a lot of work reporting unsuccessful experiences in applying so-cial recommender systems in academia. In this subsection, we present some keyfindings from both positive and negative experiences in applying social recom-mender systems to seek deeper understanding and further development of socialrecommendation.


3.4.1 Positive Experiences in Social Recommendation

In the physical world, users often seek recommendations from their friends; wehave similar observations in the online worlds. For example, 66% of people on so-cial sites have asked friends or followers to help them make a decision and 88% oflinks that 14-24 year olds clicked were sent to them by a friend and 78% of con-sumers trust peer recommendations over ads and Google SERPs10. The intuitionof social recommendation makes sense and that is why people believe social recom-mendation can potentially improve recommendation performance. There have beenmany successful social recommender systems proposed in recent years in academiaas reviewed in the above subsections, and we summarize some key findings fromthese successful experiences below.

First, online users are inherently correlated [69]. They rarely make decisionsindependently and usually seek advice from their friends before making purchasedecisions. A user’s preference is more likely to be similar to that of her socialnetwork than to those of randomly chosen users. This phenomenon is widely ob-served in many online social networks such as following relations in Twitter [120]and trust relations in Epinions [109]. This phenomenon can be explained by well-studied social theories such as homophily and social influence, which supports theutility of social recommendation. Furthermore, there is little overlap between auser’s social network and her similar users [19]. Connected users provide differ-ent information from similar users for recommendation, which can be exploited toimprove the quality of recommendations [49].

Second, to create good quality recommendations, traditional recommender sys-tems need enough historical ratings from each user. Since the rating matrix isusually very sparse due to most users rating few of the millions of items, twousers don’t have enough of the number of items rated in common required byuser similarity metrics to compute similarity. Therefore, the system is forced tochoose neighbors in the small portion of comparable users and is probably going tomiss other non-comparable but relevant users [75]. Most of these systems are notable to generate accurate recommendations for users with few or no ratings. So-cial recommendation can make recommendations as long as the user is connectedto a large enough component of the social network [50], hence social recommen-dation can significantly reduce cold-start users. For example, in [75], traditionalrecommender systems totally fail for new users, however, by considering ratingsof trusted users, social recommendation achieves a very small error and is able toproduce a recommendation for almost 17% of the users.

Finally, on a very sparse dataset that contains a large portion of cold startusers and of items rated by just one user, coverage becomes an important issuesince many of the ratings become hardly predictable [64]. Coverage refers to thefraction of ratings for which, after being hidden, the recommender systems are ableto produce a predicted rating. By propagating trust, it is possible to reach moreusers; hence, to compute a predicted trust score among them and to count themas neighbors, social recommendation can improve the coverage of recommendationespecially for new items. For example, in [31], traditional recommender systemscannot be applied to new locations; however, social recommendation can achieve

10 http://www.firebellymarketing.com/2009/12/social-search-statistics-fromses-chicago.html


more than 20% recommendation accuracy for new locations when exploiting socialrelations.

In summary, the key findings from successful experiences are - (1) social infor-mation contains complementary information and results from social network anal-ysis provide the necessary technical support for social recommendation; (2) users’opinions and tastes can be propagated via social networks, which can reduce thesize of cold-start users; (3) social recommendation can significantly improve thecoverage of recommendation.

3.4.2 Negative Experiences in Social Recommendation

Due to the potential value of social information in recommender systems, so-cial recommendation is aggressively pursued in industry. However, a recent reportshows that the startup ideas of social recommendations consistently fail [92]. IBM’sBlack Friday report says Twitter delivered 0 percent of referral traffic and Face-book sent just 0.68 percent [48]. Even in academia, unsuccessful attempts at socialrecommender systems are also reported [59,15,30]. In this subsection, we summa-rize some key points from negative experiences in applying social recommendersystems.

First, the low cost of formation of connections allows one to have an inordi-nate number of friends in the online world. For example, a Facebook user has 130friends on average, and an average Twitter user has 126 followers. Research byRobin Dunbar indicates that 100 to 150 is the approximate natural group size inwhich everyone can really know each other because our minds are not designed toallow us to have more than a very limited number of people in our social world.The emotional and psychological investments that a close relationship requires areconsiderable, and the emotional capital we have available is limited [24]. Since thesocial network comprises of valuable friends, casual friends and event friends, usersare not necessarily all that similar and social relations mixed with useful and noiseconnections may introduce negative information into recommender systems [92].For example, social recommender systems simply using all available relations per-form worse than traditional recommender systems [6,110].

Second, the available social relations are extremely sparse, and the distribu-tion of the number of social relations follows a power-law-like distribution [84],suggesting that a small number of users specify many social relations while a largeproportion of users specify a few relations. Users with many social relations arelikely to be active users [4] and they are likely to have many ratings, while userswith fewer ratings are likely to also have fewer connections [109]. For users withenough ratings, traditional recommender systems already perform well, while forusers with fewer ratings, they are likely to also have fewer social relations andsocial recommender systems cannot help much. For example, social recommendersystems in [59,15,30] gain a little or even no improvement compared to traditionalrecommender systems.

Finally, most existing successful social recommender systems use trust relationsand recommend items to a user from her trusted users. Trust plays a central role inexchanging relationships involving unknown risk [32], which provides informationabout with whom we should share information and from whom we should acceptinformation [35]. The role of trust is especially critical in some online communitiessuch as e-commerce sites and product review sites, which has been described as


the “wild wild west” of the 21st century [76]. These successful social recommendersystems assume that trust relations are formed when users have similar opinionsto similar products. With this assumption, users are likely to seek recommenda-tions from their trusted friends and social recommender systems are likely to besuccessful. However, trust is a complex concept and has different interpretationsin different contexts [21,27,2]. A user in the context of product review sites suchas Epinions will trust the users if she agrees with their opinions about products,while a user in the context of P2P networks will trust others because of theirreliability. Different interpretations of trust may result in different solutions forsocial recommendation, which suggests that more nuanced approaches are neededto exploit trust relations for recommendation. Trust relations are not necessar-ily equivalent to other types of social relations. For example, following a user inTwitter does not indicate one’s trust in the user. The success of exploiting trustrelations for recommendation may not be applied to other relations, and differenttypes of social relations may have very different impacts on social recommendersystems [125].

In summary, key findings from negative experiences in applying social rec-ommender systems include - (1) social relations are too noisy and may have anegative impact on recommender systems; (2) cold-start users are likely to alsohave few or no social relations and it is difficult for social recommender systems toimprove recommendation performance for these users; and (3) different types ofsocial relations have different effects on social recommender systems, and success-ful experiences in exploiting trust relations may not be applicable to other typesof relations.

3.5 Evaluation

In this subsection, we discuss the evaluation of social recommender systems, fo-cusing on available datasets and evaluation metrics for social recommendation.

3.5.1 Datasets

There are benchmark datasets to evaluate traditional recommender systems suchas Netflix data 11 and MovieLens data 12. Compared to traditional recommenda-tion, social recommendation has recently emerged with the development of socialmedia, and there are no agreed benchmark datasets. However, there are datasetspublicly available for the purpose of research; here is a list of several representativeones.

Epinions03 13: This dataset was collected by Paolo Massa in a 5-week crawl(November/December 2003) from the Epinions.com [73]. In Epinions, people canwrite reviews for various products with ratings, and also they can add members totheir trust networks or “Circle of Trust”. This dataset provides user-item ratinginformation and user-user trust networks. Note that the trust network here isdirected.

11 http://en.wikipedia.org/wiki/Netflix Prize12 http://www.grouplens.org/node/7313 http://www.trustlet.org/wiki/Downloaded Epinions dataset


Flixster 14: This dataset was crawled from the Flixster website [50], which is asocial networking service. In Flixster, users can rate movies and can also add someusers to their friend list, establishing social networks. Flixster provides user-itemrating information and user-user friendship networks. Note that social relations inFlixster are undirected.

Ciao 15: This dataset was collected by authors in [110,111]. Ciao is a productreview website where users can rate and write reviews for various products, andthey can also establish social relations with others. In addition to rating and socialinformation, Ciao provides extra contextual information including temporal infor-mation about when ratings are provided, the category information of products,and information about the reviews such as the content and helpfulness votes.

Epinions11 16: This dataset was collected during 2011 and is also from Epin-ions [110,111]. In addition to information in Epinion03, this dataset includes richerinformation for social recommendation, including temporal information for bothrating and social information, categories of products, information about reviews,and distrust information, which allows advanced research about social recommen-dation. For example, the temporal information about when users establish trustrelations can be used to study the evolution of social relations in social recommen-dation.

Note that some social recommendation papers use other datasets such as Epin-ions02 [128], FilmTrust [34], and Douban [68]. We do not give details to thembecause these datasets are publicly unavailable or similar to datasets listed above.

3.5.2 Evaluation Metrics

Although the inputs of traditional recommendation and social recommendationare different, their outputs are the same, i.e., the predicted values for unknownratings. Therefore, metrics that evaluate traditional recommender systems can alsobe applied to evaluate social recommender systems.

To evaluate recommender systems, the data is usually divided into two parts-the training set K (known ratings) and the testing set U (unknown ratings). Rec-ommender systems will be trained based on K, and the quality of recommendationwill be evaluated in U . Different evaluation metrics are proposed to evaluate thequality of recommendation from different perspectives, such as prediction accu-racy, ranking accuracy, diversity and novelty, and coverage. Prediction accuracyand ranking accuracy are two widely adopted metrics.

Prediction Accuracy: Prediction accuracy measures the closeness of pre-dicted ratings to the true ratings. Two widely used metrics in this category areMean Absolute Error (MAE) and Root Mean Squared Error (RMSE).

The metric RMSE is defined as

RMSE =

√

∑

(ui,vj)∈U (Rij − Rij)2

|U| , (28)

where |U| is the size of U and Rij is the predicted rating from ui to vj .

14 http://www.cs.ubc.ca/ jamalim/datasets/15 http://www.public.asu.edu/ jtang20/datasetcode/truststudy.htm16 http://www.public.asu.edu/ jtang20/datasetcode/truststudy.htm


The metric MAE is defined as

MAE =1

|U|∑

(ui,vj)∈U

|Rij − Rij |. (29)

A smaller RMSE or MAE value means better performance, and due to theirsimplicity, RMSE and MAE are widely used in the evaluation of recommender sys-tems. Note that previous work demonstrated that small improvement in RMSEor MAE terms can have a significant impact on the quality of the top-few recom-mendation [57].

Ranking Accuracy: Ranking accuracy evaluates how many recommendeditems are purchased by the user. Precision and recall are two popular metricsin this category. Recall captures how many of the acquired items are recom-mended, while precision captures how many recommended items are acquired,for example, Prec@N is used to indicate how many top-N recommended itemsare acquired. Long recommendation lists typically improve recall, while reducingprecision. Therefore F-score is a metric combining them, and it is less dependenton the length of the recommendation list.

Another popular metric is Discount Cumulative Gain (DCG), which is definedas

DCG =1

|u|∑

ui∈u

|L|∑

j=1

Rij

max (1, logb j)(30)

where L is the ranked list of recommended items.In this survey, we do not give a general comparison of existing social recom-

mender systems and the reasons are three-fold. First, although some datasets arepublicly available, there are still no agreed benchmark datasets for social recom-mendation. Second, social recommendation is a complex problem and differentalgorithms are designed to capture its different aspects. For example, algorithmsin [117,118] are for controversial items, while MoleTrust is for cold-start users.Third, some algorithms may require additional sources to capture social infor-mation more effectively, for example, mTrust needs the category information ofitems [110], therefore, it is difficult to seek a fair comparison for all systems. How-ever, there are some common questions to answer when accessing a recommendersystem including (1) which datasets we are going to adopt; (2) which metrics wewill use; (3) what kinds of items we want to recommend; and (4) which kinds ofusers we want to recommend to. To help readers further understand the assessmentprocess, we summarize evaluations of representative social recommender systemsw.r.t. these questions in Figure 2.

4 Research Directions of Social Recommendation

Since the performance boost varies from domain to domain, social recommendationis still in the early stages of development and an active area of exploration. In thissubsection, we discuss several research directions that can potentially improvethe capabilities of social recommender systems and make social recommendationapplicable to an even broader range of applications.


Algorithms Datasets Metrics Items Users

D1 D2 D3 D4 D5 D6 D7 M1 M2 M3 M4 M5 I1 I2 U1 U2

SBWM[117,118] √ √ √ √ √ √

TidalTrust[34] √ √ √ √

MoleTrust[73-75] √ √ √ √ √ √

TrustWalker[49] √ √ √ √ √ √ √

SoRec[67] √ √ √ √ √

LOCABAL[109] √ √ √ √ √ √

STE[65] √ √ √ √ √ √

mTrust[110] √ √

SocialMF[50] √ √ √ √ √

SoReg[68] √ √ √ √ √ √

Fig. 2 Summarization of Evaluations of Representative Social RecommenderSystems. Note that the meanings of symbols in the table are stated asD1(Epinions03), D2(Flixster), D3(Ciao), D4(Epinions11), D5(Epinions02), D6(FilmTrust), D7(Douban),M1(RMSE), M2(MAE), M3(Coverage), M4(F-Measure), M5(Others),I1(Controversial Items), I2(Other Items), and U1(Cold-start Users), U2(Other Users).

4.1 The Heterogeneity of Social Networks

Most existing social recommender systems treat a user’s connections homoge-neously. However, connections in online social networks are intrinsically hetero-geneous and are a composite of various types of relations [113,106,110]. Figure 3illustrates an example of u1’s social relations with u2, u3, . . . , u9. The user u1

may treat her social relations differently in different domains. For example, u1

may seek suggestions about “Sports” from u2, u3, but ask for recommendationabout “Electronics” from u4, u5. In [110], the authors found that people placetrust differently to users in different domains. For example, ui might trust uj in“Sports” but not trust uj in “Electronics” at all. For different sets of items, ex-ploiting different types of social relations can potentially benefit existing socialrecommender systems [110].

4.2 Weak Dependence Connections

Most existing model-based social recommender systems solely make use of a user’sstrong dependence connections, i.e., direct connections, which underestimates thediversity of users’ opinions and tastes [92]. If users in the physical world onlyhad strong dependence connections, then life would be pretty boring since strongdependence connections indicate strong similarities. Actually, users can establishweak dependence connections with others in social networks when they are notdirectly connected. Weak dependency connections can provide important context


!

" "

#

$

$

#

%

&

'

'

%

&

!

(

(

Fig. 3 An illustration of u1’s social connections

information about users’ interests, and are proven to be useful in job hunting [39],the diffusion of ideas [40], knowledge transfer [61] and relational learning [113],while rarely used in recommendation.

Identifying weak dependence connections for recommendation is an interestingdirection to investigate. One possible way to find weak dependence connectionsis to exploit users’ geo-locations. For example, [98] notices that users geographi-cally close are likely to share similar interests, and that users geographically closeare likely to visit similar locations [31]. Another possible way is to detect groupsin social networks. Connected users in online social networks form groups wherethere are more connections among users within groups than among those betweengroups [85,29]. For example, in Figure 3, u1, u2, u3, u10, u11, u12 form a groupwhile u1, u3, u5, u13, u14, u15 form another group. According to social correlationtheories, similar users interact at higher rates than dissimilar ones; thus, users inthe same group are likely to share similar preferences, establishing weak depen-dency connections when they are not directly connected [113].

4.3 User Segmentation

In traditional recommender systems, for a given user, ratings of users most similarto her are aggregated to predict a missing rating. When involved in social infor-mation, besides being similar, users are socially connected. A user’s most similar


Users Items

III

II I

IV

1

2

?

Fig. 4 User segmentation for social recommendation, where “I”, “II”, “III” and “IV” denoteconnected and non-similar users, connected and similar users, non-connected and similar users,and non-connected and non-similar users respectively, while “1” and “2” denote cold-startitems and normal items, respectively.

users usually have little overlap with her connected users [19]. Therefore, userscan be segmented into four groups as illustrated in Figure 4 - I: connected andnon-similar users; II: connected and similar users; III: non-connected and similarusers; and IV: non-connected and non-similar users. According to the number ofattracted ratings, items can be segmented into cold-start items and normal items.Different types of users may contribute differently for different types of items. Forexample, connected users can improve the recommendation accuracy of cold-startlocations [31], while similar users are important to recommend normal items [73].Therefore, microcosmic investigations of users and items may give us a deeperunderstanding of the role of social networks and can potentially improve recom-mendation performance.

4.4 Temporal Information

Customer preferences for products drift over time. For example, people interestedin “Electronics” at time t may shift their preferences to “Sports” at time t + 1.Temporal information is an important factor in recommender systems and thereare traditional recommender systems that consider temporal information [23,58].Temporal dynamics in data can have a more significant impact on accuracy thandesigning more complex learning algorithms [58]. Exploiting temporal informa-tion for recommender systems is still challenging due to the complexity of users’temporal patterns [58].

Social relations also change over time. For example, new social relations areadded, while existing social relations become inactive or are deleted. The changesof both ratings and social relations further exacerbate the difficulty of exploitingtemporal information for social recommendation. A preliminary study of the im-pact of the changes in both ratings and trust relations on recommender systemsdemonstrates that temporal information can benefit social recommendation [111].


4.5 Negative Relations

Currently most existing social recommender systems use positive relations suchas friendships and trust relations. However, in social media, users also specifynegative relations such as distrust and dislike. Authors in [1] found that negativerelations are even more important than positive relations, revealing the importanceof negative relations for social recommendation. There are several works exploit-ing distrust [66,117] in social recommender systems. They treat trust and distrustseparately, and simply use distrust in an opposite way to trust such as filteringdistrusted users or considering distrust relations as negative weights. However,trust and distrust are shaped by different dimensions of trustworthiness, and trustaffects behavior intentions differently from distrust [16]. Furthermore, distrust re-lations are not independent of trust relations [118]. A deeper understanding ofnegative relations and the correlations to positive relations can help us develop ef-ficient social recommender systems by exploiting both positive and negative socialrelations.

4.6 Cross Media Data

A user generally has multiple accounts in social media. For example, a user whohas an account in Epinions might also have an account in eBay. A new user onone website might have existed on another website for a long time. For example,a user has already specified her interests in Epinions and has also written manyreviews about items. When the user registers at eBay for the first time as a cold-start user, data about the user in Epinions can help eBay solve the cold-startproblem and accurately recommend items to the user. Integrating networks frommultiple websites can bring about a huge impact on social recommender systemsand provide an efficient and effective way to solve the cold-start problem. The firstdifficulty of integrating data is connecting corresponding users across websitesand there is recent work proposed to tackle this mapping problem [83,127,63].The study of the mapping problem makes integration of cross-media data forsocial recommendation possible and brings about new opportunities for socialrecommender systems.

5 Conclusion

Social recommendation has attracted broad attention from both academia andindustry, and many social recommender systems have been proposed in recentyears. In this paper, we first give a narrow definition and a broad definition of so-cial recommendation to cover most existing definitions of social recommendationin literature, and discuss the unique feature of social recommender systems as wellas its implications. We classify current social recommender systems into memory-based social recommender systems and model-based social recommender systemsaccording to the basic models chosen to build the systems, and then present a re-view of representative systems for each category. We also discuss some key findingsfrom positive and negative experiences in applying social recommender systems.Social recommendation is still in the early stages of development and needs further


improvement. Finally we present research directions that can potentially improveperformance of social recommender systems including exploiting the heterogeneityof social networks and weak dependence connections, microcosmic investigation ofusers and items, considering temporal information in rating and social information,understanding the role of negative relations, and integrating cross-media data.

References

1. Abbassi, Z., Aperjis, C., Huberman, B.A.: Friends versus the crowd: tradeoffs and dy-namics. HP Report (2013)

2. Adali, S.: Modeling Trust Context in Networks. Springer (2013)3. Adomavicius, G., Tuzhilin, A.: Toward the next generation of recommender systems: A

survey of the state-of-the-art and possible extensions. IEEE Transactions on Knowledgeand Data Engineering, 17(6), 734–749 (2005)

4. Agarwal, N., Liu, H., Tang, L., Yu, P.: Identifying the influential bloggers in a community.In: Proceedings of the international conference on Web search and web data mining, pp.207–218. ACM (2008)

5. Agarwal, V., Bharadwaj, K.: A collaborative filtering framework for friends recommenda-tion in social networks based on interaction intensity and adaptive user similarity. SocialNetwork Analysis and Mining pp. 1–21 (2012)

6. Au Yeung, C., Iwata, T.: Strength of social influence in trust networks in product reviewsites. In: Proceedings of the fourth ACM international conference on Web search anddata mining, pp. 495–504. ACM (2011)

7. Baeza-Yates, R., Ribeiro-Neto, B., et al.: Modern information retrieval, vol. 463. ACMpress New York. (1999)

8. Balabanovic, M., Shoham, Y.: Fab: content-based, collaborative recommendation. Com-munications of the ACM 40(3), 66–72 (1997)

9. Belkin, N.J., Croft, W.B.: Information filtering and information retrieval: two sides ofthe same coin? Communications of the ACM 35(12), 29–38 (1992)

10. Breese, J.S., Heckerman, D., Kadie, C.: Empirical analysis of predictive algorithms forcollaborative filtering. In: Proceedings of the Fourteenth conference on Uncertainty inartificial intelligence, pp. 43–52. (1998)

11. Chee, S.H.S., Han, J., Wang, K.: Rectree: An efficient collaborative filtering method. In:Data Warehousing and Knowledge Discovery, pp. 141–151. Springer (2001)

12. Chen, B., Guo, J., Tseng, B., Yang, J.: User reputation in a comment rating environ-ment. In: Proceedings of the 17th ACM SIGKDD international conference on Knowledgediscovery and data mining, pp. 159–167. ACM (2011)

13. Chen, J., Geyer, W., Dugan, C., Muller, M., Guy, I.: Make new friends, but keep theold: recommending people on social networking sites. In: Proceedings of the SIGCHIConference on Human Factors in Computing Systems, pp. 201–210. ACM (2009)

14. Chen, W.Y., Chu, J.C., Luan, J., Bai, H., Wang, Y., Chang, E.Y.: Collaborative filteringfor orkut communities: discovery of user latent behavior. In: Proceedings of the 18thinternational conference on World wide web, pp. 681–690. ACM (2009)

15. Cho, E., Myers, S.A., Leskovec, J.: Friendship and mobility: user movement in location-based social networks. In: Proceedings of the 17th ACM SIGKDD international confer-ence on Knowledge discovery and data mining, pp. 1082–1090. ACM (2011)

16. Cho, J.: The mechanism of trust and distrust formation and their relational outcomes.Journal of Retailing 82(1), 25–35 (2006)

17. Chowdhury, G.: Introduction to modern information retrieval. Facet publishing (2010)18. Claypool, M., Gokhale, A., Miranda, T., Murnikov, P., Netes, D., Sartin, M.: Combining

content-based and collaborative filters in an online newspaper. In: Proceedings of ACMSIGIR workshop on recommender systems, vol. 60. Citeseer (1999)

19. Crandall, D., Cosley, D., Huttenlocher, D., Kleinberg, J., Suri, S.: Feedback effects be-tween similarity and social influence in online communities. In: Proceedings of the 14thACM SIGKDD international conference on Knowledge discovery and data mining, pp.160–168. ACM (2008)

20. Davis, D., Lichtenwalter, R., Chawla, N.V.: Supervised methods for multi-relational linkprediction. Social Network Analysis and Mining pp. 1–15 (2013)


21. Dellarocas, C., Zhang, X.M., Awad, N.F.: Exploring the value of online product reviewsin forecasting sales: The case of motion pictures. Journal of Interactive marketing 21(4),23–45 (2007)

22. Deshpande, M., Karypis, G.: Item-based top-n recommendation algorithms. ACM Trans-actions on Information Systems 22(1), 143–177 (2004)

23. Ding, Y., Li, X.: Time weight collaborative filtering. In: Proceedings of the 14th ACMinternational conference on Information and knowledge management, pp. 485–492. ACM(2005)

24. Dunbar, R.: How many friends does one person need? Faber & Faber (2010)25. Dunlavy, D., Kolda, T., Acar, E.: Temporal link prediction using matrix and tensor

factorizations. ACM Transactions on Knowledge Discovery from Data (TKDD) 5(2), 10(2011)

26. Ellenberg, J.: This psychologist might outsmart the math brains competing for the netflixprize. Wired Magazine, March pp. 114–122 (2008)

27. Falcone, R., Castelfranchi, C.: Transitivity in trust a discussed property. Citeseer (2010)28. Fang, Y., Si, L.: Matrix co-factorization for recommendation with rich side information

and implicit feedback. In: Proceedings of the 2nd International Workshop on InformationHeterogeneity and Fusion in Recommender Systems, pp. 65–69. ACM (2011)

29. Fortunato, S.: Community detection in graphs. Physics Reports 486(3), 75–174 (2010)30. Gao, H., Tang, J., Liu, H.: Exploring social-historical ties on location-based social net-

works. In: Proceedings of the 6th international AAAI conference on weblogs and socialmedia (2012)

31. Gao, H., Tang, J., Liu, H.: gscorr: Modeling geo-social correlations for new check-ins onlocation-based social networks. In: Proceedings of the 21st ACM international conferenceon Information and knowledge management, pp. 1582–1586. ACM (2012)

32. Gefen, D., Karahanna, E., Straub, D.: Trust and tam in online shopping: An integratedmodel. Mis Quarterly pp. 51–90 (2003)

33. Golbeck, J.: Generating predictive movie recommendations from trust in social networks.Trust Management pp. 93–104 (2006)

34. Golbeck, J.: Generating predictive movie recommendations from trust in social networks.Springer (2006)

35. Golbeck, J.: Trust and nuanced profile similarity in online social networks. ACM Trans-actions on the Web 3(4), 1–33 (2009)

36. Goldberg, D., Nichols, D., Oki, B.M., Terry, D.: Using collaborative filtering to weave aninformation tapestry. Communications of the ACM 35(12), 61–70 (1992)

37. Goldberg, K., Roeder, T., Gupta, D., Perkins, C.: Eigentaste: A constant time collabo-rative filtering algorithm. Information Retrieval 4(2), 133–151 (2001)

38. Good, N., Schafer, J.B., Konstan, J.A., Borchers, A., Sarwar, B., Herlocker, J., Riedl, J.:Combining collaborative filtering with personal agents for better recommendations. In:Proceedings of the National Conference on Artificial Intelligence, pp. 439–446. (1999)

39. Granovetter, M.: The strength of weak ties. American Journal of Socialogy 78(6), 1360–1380 (1973)

40. Granovetter, M.: The strength of weak ties: A network theory revisited. Sociologicaltheory 1(1), 201–233 (1983)

41. Guy, I., Carmel, D.: Social recommender systems. In: Proceedings of the 20th interna-tional conference companion on World wide web, pp. 283–284. ACM (2011)

42. Guy, I., Jacovi, M., Perer, A., Ronen, I., Uziel, E.: Same places, same things, same people?:mining user similarity on social media. In: Proceedings of the 2010 ACM conference onComputer supported cooperative work, pp. 41–50. ACM (2010)

43. Guy, I., Jacovi, M., Shahar, E., Meshulam, N., Soroka, V., Farrell, S.: Harvesting withsonar: the value of aggregating social network information. In: Proceedings of the SIGCHIConference on Human Factors in Computing Systems, pp. 1017–1026. ACM (2008)

44. Herlocker, J., Konstan, J., Borchers, A., Riedl, J.: An algorithmic framework for per-forming collaborative filtering. In: Proceedings of the 22nd annual international ACMSIGIR conference on Research and development in information retrieval, pp. 230–237.ACM (1999)

45. Hofmann, T.: Latent semantic models for collaborative filtering. ACM Transactions onInformation Systems (TOIS) 22(1), 89–115 (2004)

46. Hong, L., Doumith, A.S., Davison, B.D.: Co-factorization machines: modeling user in-terests and predicting individual decisions in twitter. In: Proceedings of the sixth ACMinternational conference on Web search and data mining, pp. 557–566. ACM (2013)


47. Huang, Z., Zeng, D., Chen, H.: A link analysis approach to recommendation under sparsedata. In: Proc. 2004 Americas Conf. Information Systems (2004)

48. IBM: Ibm’s black friday report says twitter delivered 0 percent of referral traffic andfacebook sent just 0.68 percent. In: https://strme.wordpress.com/2012/11/27/ibms-black-friday-report-says-twitter-delivered-0-percent-of-referral-traffic-and-facebook-sent-just-0-68-percent/ (2012)

49. Jamali, M., Ester, M.: Trustwalker: a random walk model for combining trust-based anditem-based recommendation. In: Proceedings of the 15th ACM SIGKDD internationalconference on Knowledge discovery and data mining, pp. 397–406. ACM (2009)

50. Jamali, M., Ester, M.: A matrix factorization technique with trust propagation for rec-ommendation in social networks. In: Proceedings of the fourth ACM conference onRecommender systems, pp. 135–142. ACM (2010)

51. Jannach, D., Zanker, M., Felfernig, A., Friedrich, G.: Recommender systems: an intro-duction. Cambridge University Press (2010)

52. Jiang, M., Cui, P., Liu, R., Yang, Q., Wang, F., Zhu, W., Yang, S.: Social contextual rec-ommendation. In: Proceedings of the 22th ACM international conference on Informationand knowledge management. ACM (2012)

53. Karypis, G.: Evaluation of item-based top-n recommendation algorithms. In: Proceedingsof the tenth international conference on Information and knowledge management, pp.247–254. ACM (2001)

54. Kautz, H., Selman, B., Shah, M.: Referral web: combining social networks and collabo-rative filtering. Communications of the ACM 40(3), 63–65 (1997)

55. King, I., Lyu, M.R., Ma, H.: Introduction to social recommendation. In: Proceedings ofthe 19th international conference on World wide web, pp. 1355–1356. ACM (2010)

56. Kleinberg, J.: Authoritative sources in a hyperlinked environment. Journal of the ACM46(5), 604–632 (1999)

57. Koren, Y.: Factorization meets the neighborhood: a multifaceted collaborative filteringmodel. In: Proceeding of the 14th ACM SIGKDD international conference on Knowledgediscovery and data mining, pp. 426–434. ACM (2008)

58. Koren, Y.: Collaborative filtering with temporal dynamics. In: Proceedings of the 15thACM SIGKDD international conference on Knowledge discovery and data mining, pp.447–456. ACM (2009)

59. Leskovec, J., Adamic, L.A., Huberman, B.A.: The dynamics of viral marketing. ACMTransactions on the Web 1(1), 5 (2007)

60. Leskovec, J., Huttenlocher, D., Kleinberg, J.: Predicting positive and negative links inonline social networks. In: Proceedings of the 19th international conference on Worldwide web (2010)

61. Levin, D.Z., Cross, R.: The strength of weak ties you can trust: The mediating role oftrust in effective knowledge transfer. Management science 50(11), 1477–1490 (2004)

62. Li, Y., Hu, J., Zhai, C., Chen, Y.: Improving one-class collaborative filtering by incorpo-rating rich user information. In: Proceedings of the 19th ACM international conferenceon Information and knowledge management, pp. 959–968. ACM (2010)

63. Liu, J., Zhang, F., Song, X., Song, Y.I., Lin, C.Y., Hon, H.W.: What’s in a name?:an unsupervised approach to link users across communities. In: Proceedings of the sixthACM international conference on Web search and data mining, pp. 495–504. ACM (2013)

64. Lu, L., Medo, M., Yeung, C.H., Zhang, Y.C., Zhang, Z.K., Zhou, T.: Recommendersystems. Physics Reports (2012)

65. Ma, H., King, I., Lyu, M.R.: Learning to recommend with social trust ensemble. In: Pro-ceedings of the 32nd international ACM SIGIR conference on Research and developmentin information retrieval, pp. 203–210. ACM (2009)

66. Ma, H., Lyu, M.R., King, I.: Learning to recommend with trust and distrust relationships.In: Proceedings of the third ACM conference on Recommender systems, pp. 189–196.ACM (2009)

67. Ma, H., Yang, H., Lyu, M., King, I.: Sorec: social recommendation using probabilisticmatrix factorization. In: Proceeding of the 17th ACM conference on Information andknowledge management, pp. 931–940. ACM (2008)

68. Ma, H., Zhou, D., Liu, C., Lyu, M., King, I.: Recommender systems with social regular-ization. In: Proceedings of the fourth ACM international conference on Web search anddata mining, pp. 287–296. ACM (2011)

69. Ma, H., Zhou, T.C., Lyu, M.R., King, I.: Improving recommender systems by incorpo-rating social contextual information. ACM Transactions on Information Systems 29(2),9 (2011)


70. Ma, N., Lim, E., Nguyen, V., Sun, A., Liu, H.: Trust relationship prediction using onlineproduct review data. In: Proceeding of the 1st ACM international workshop on Complexnetworks meet information & knowledge management, pp. 47–54. ACM (2009)

71. Marsden, P., Friedkin, N.: Network studies of social influence. Sociological Methods andResearch 22(1), 127–151 (1993)

72. Massa, P.: A survey of trust use and modeling in real online systems. Trust in E-services:Technologies, Practices and Challenges (2007)

73. Massa, P., Avesani, P.: Trust-aware collaborative filtering for recommender systems. In:On the Move to Meaningful Internet Systems 2004: CoopIS, DOA, and ODBASE, pp.492–508. Springer (2004)

74. Massa, P., Avesani, P.: Controversial users demand local trust metrics: An experimen-tal study on epinions. com community. In: Proceedings of the National Conference onArtificial Intelligence, vol. 20, p. 121. Menlo Park, CA; Cambridge, MA; London; AAAIPress; MIT Press; 1999 (2005)

75. Massa, P., Avesani, P.: Trust-aware recommender systems. In: Proceedings of the 2007ACM conference on Recommender systems, pp. 17–24. ACM (2007)

76. McKnight, D., Choudhury, V., Kacmar, C.: Developing and validating trust measuresfor e-commerce: An integrative typology. Information systems research 13(3), 334–359(2003)

77. McPherson, M., Smith-Lovin, L., Cook, J.: Birds of a feather: Homophily in social net-works. Annual review of sociology pp. 415–444 (2001)

78. Mei, T., Yang, B., Hua, X.S., Yang, L., Yang, S.Q., Li, S.: Videoreach: an online videorecommendation system. In: Proceedings of the 30th annual international ACM SIGIRconference on Research and development in information retrieval, pp. 767–768. ACM(2007)

79. Menon, A., Elkan, C.: Link prediction via matrix factorization. Machine Learning andKnowledge Discovery in Databases pp. 437–452 (2011)

80. Menon, A.K., Elkan, C.: Link prediction via matrix factorization. In: Machine Learningand Knowledge Discovery in Databases, pp. 437–452. Springer (2011)

81. Miyahara, K., Pazzani, M.J.: Collaborative filtering with the simple bayesian classifier.In: PRICAI 2000 Topics in Artificial Intelligence, pp. 679–689. Springer (2000)

82. Mooney, R.J., Bennett, P.N., Roy, L.: Book recommending using text categorization withextracted information. In: Proc. Recommender Systems Papers from 1998 Workshop,Technical Report WS-98-08 (1998)

83. Narayanan, A., Shmatikov, V.: De-anonymizing social networks. In: Security and Privacy,2009 30th IEEE Symposium on, pp. 173–187. IEEE (2009)

84. Newman, M.E.: Power laws, pareto distributions and zipf’s law. Contemporary physics46(5), 323–351 (2005)

85. Newman, M.E.J.: Power laws, pareto distributions and zipf’s law. Contemporary Physics46, 323–351 (2005)

86. Noel, J., Sanner, S., Tran, K.N., Christen, P., Xie, L., Bonilla, E.V., Abbasnejad, E.,Della Penna, N.: New objective functions for social collaborative filtering. In: Proceedingsof the 21st international conference on World Wide Web, pp. 859–868. ACM (2012)

87. Page, L., Brin, S., Motwani, R., Winograd, T.: The pagerank citation ranking: Bringingorder to the web. Stanford InfoLab(1999)

88. Pan, R., Zhou, Y., Cao, B., Liu, N.N., Lukose, R., Scholz, M., Yang, Q.: One-class collab-orative filtering. In: Eighth IEEE International Conference on Data Mining, pp. 502–511.IEEE (2008)

89. Paterek, A.: Improving regularized singular value decomposition for collaborative filter-ing. In: Proceedings of KDD cup and workshop, vol. 2007, pp. 5–8 (2007)

90. Pazzani, M., Billsus, D.: Learning and revising user profiles: The identification of inter-esting web sites. Machine learning 27(3), 313–331 (1997)

91. Pazzani, M.J.: A framework for collaborative, content-based and demographic filtering.Artificial Intelligence Review 13(5-6), 393–408 (1999)

92. Quora: Why does the startup idea of social recommendations consistentlyfail? In: http://www.quora.com/Why-does-the-startup-idea-of-social-recommendations-consistently-fail (2012)

93. Raghavan, S., Gunasekar, S., Ghosh, J.: Review quality aware collaborative filtering. In:Proceedings of the sixth ACM conference on Recommender systems, pp. 123–130. ACM(2012)


94. Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P., Riedl, J.: Grouplens: an open archi-tecture for collaborative filtering of netnews. In: Proceedings of the 1994 ACM conferenceon Computer supported cooperative work, pp. 175–186. ACM (1994)

95. Ricci, F., Rokach, L., Shapira, B.: Introduction to recommender systems handbook. Rec-ommender Systems Handbook pp. 1–35 (2011)

96. Salakhutdinov, R., Mnih, A.: Probabilistic matrix factorization. Advances in neuralinformation processing systems 20, 1257–1264 (2008)

97. Sarwar, B., Karypis, G., Konstan, J., Riedl, J.: Item-based collaborative filtering recom-mendation algorithms. In: Proceedings of the 10th international conference on WorldWide Web, pp. 285–295. ACM (2001)

98. Scellato, S., Mascolo, C., Musolesi, M., Latora, V.: Distance matters: geo-social metricsfor online social networks. Proceedings of WOSN 10 (2010)

99. Schafer, J.B., Konstan, J.A., Riedl, J.: E-commerce recommendation applications. Datamining and knowledge discovery 5(1), 115–153 (2001)

100. Scott, J.: Social network analysis: developments, advances, and prospects. Social networkanalysis and mining 1(1), 21–26 (2011)

101. Scott, J.: Social network analysis. SAGE Publications Limited (2012)102. Sigurbjornsson, B., Van Zwol, R.: Flickr tag recommendation based on collective knowl-

edge. In: Proceedings of the 17th international conference on World Wide Web, pp.327–336. ACM (2008)

103. Sinha, R., Swearingen, K.: Comparing recommendations made by online systems andfriends. In: Proceedings of the Delos-NSF workshop on personalization and recommendersystems in digital libraries, vol. 106. Dublin, Ireland (2001)

104. Soboroff, I., Nicholas, C.: Combining content and collaboration in text filtering. In:Proc. Intl Joint Conf. Artificial Intelligence Workshop: Machine Learning for InformationFiltering (1999)

105. Su, X., Khoshgoftaar, T.: A survey of collaborative filtering techniques. Advances inArtificial Intelligence 2009, 4 (2009)

106. Sun, Y., Han, J.: Mining heterogeneous information networks: Principles and methodolo-gies. Synthesis Lectures on Data Mining and Knowledge Discovery 3(2), 1–159 (2012)

107. Symeonidis, P., Tiakas, E., Manolopoulos, Y.: Product recommendation and rating pre-diction based on multi-modal social networks. In: Proceedings of the fifth ACM conferenceon Recommender systems, pp. 61–68. ACM (2011)

108. Tang, J., Gao, H., Hu, X., Liu, H.: Context-aware review helpfulness rating prediction.In: RecSys (2013)

109. Tang, J., Gao, H., Hu, X., Liu, H.: Exploiting homophily effect for trust prediction. In:Proceedings of the sixth ACM international conference on Web search and data mining,pp. 53–62. ACM (2013)

110. Tang, J., Gao, H., Liu, H.: mTrust: Discerning multi-faceted trust in a connected world.In: Proceedings of the fifth ACM international conference on Web search and data mining,pp. 93–102. ACM (2012)

111. Tang, J., Gao, H., Liu, H., Das Sarma, A.: eTrust: Understanding trust evolution in anonline world. In: Proceedings of the 18th ACM SIGKDD international conference onKnowledge discovery and data mining, pp. 253–261. ACM (2012)

112. Tang, J., Hu, X., Gao, H., Liu, H.: Exploiting local and global social context for recom-mendation. In: IJCAI (2013)

113. Tang, L., Liu, H.: Relational learning via latent social dimensions. In: Proceedings of the15th ACM SIGKDD international conference on Knowledge discovery and data mining,p. 817-26 (2009)

114. Tang, L., Liu, H.: Community detection and mining in social media. Synthesis Lectureson Data Mining and Knowledge Discovery 2(1), 1–137 (2010)

115. Ungar, L.H., Foster, D.P.: Clustering methods for collaborative filtering. In: AAAI Work-shop on Recommendation Systems, 1 (1998)

116. Vasuki, V., Natarajan, N., Lu, Z., Dhillon, I.S.: Affiliation recommendation using auxil-iary networks. In: Proceedings of the fourth ACM conference on Recommender systems,pp. 103–110. ACM (2010)

117. Victor, P., Cornelis, C., De Cock, M., Teredesai, A.M.: A comparative analysis of trust-enhanced recommenders for controversial items. In: Proc. of the International AAI Con-ference on Weblogs and Social Media, pp. 342–345 (2009)

118. Victor, P., De Cock, M., Cornelis, C.: Trust and recommendations. In: RecommenderSystems Handbook, pp. 645–675. Springer (2011)


119. Wasserman, S., Faust, K.: Social network analysis: Methods and applications, vol. 8.Cambridge university press (1994)

120. Weng, J., Lim, E., Jiang, J., He, Q.: Twitterrank: finding topic-sensitive influential twit-terers. In: Proceedings of the third ACM international conference on Web search anddata mining, pp. 261–270. ACM (2010)

121. Wu, H.C., Luk, R.W.P., Wong, K.F., Kwok, K.L.: Interpreting tf-idf term weights asmaking relevance decisions. ACM Transactions on Information Systems (TOIS) 26(3),13 (2008)

122. Xiang, R., Neville, J., Rogati, M.: Modeling relationship strength in online social net-works. In: Proceedings of the 19th international conference on World wide web (2010)

123. Yang, S.H., Long, B., Smola, A., Sadagopan, N., Zheng, Z., Zha, H.: Like like alike:joint friendship and interest propagation in social networks. In: Proceedings of the 20thinternational conference on World wide web, pp. 537–546. ACM (2011)

124. Yildirim, H., Krishnamoorthy, M.S.: A random walk method for alleviating the spar-sity problem in collaborative filtering. In: Proceedings of the 2008 ACM conference onRecommender systems, pp. 131–138. ACM (2008)

125. Yuan, Q., Chen, L., Zhao, S.: Factorization vs. regularization: fusing heterogeneous socialrelationships in top-n recommendation. In: Proceedings of the fifth ACM conference onRecommender systems, pp. 245–252. ACM (2011)

126. Yuan, Q., Zhao, S., Chen, L., Liu, Y., Ding, S., Zhang, X., Zheng, W.: Augmentingcollaborative recommender by fusing explicit social relationships. In: Workshop on Rec-ommender Systems and the Social Web, Recsys 2009 (2009)

127. Zafarani, R., Liu, H.: Connecting corresponding identities across communities. In: Pro-ceedings of the 3rd International Conference on Weblogs and Social Media (ICWSM09)(2009)

128. Matthew, R., Pedro, D.: Mining knowledge-sharing sites for viral marketing. In: Pro-ceedings of the eighth ACM SIGKDD international conference on Knowledge discoveryand data mining (2002)

social recommendation: a review - ecology labfaculty.cs.tamu.edu/xiahu/papers/snam13.pdf · social...

Documents